基于改进的注意力机制与纹理结构分层的生成对抗网络图像修复方法  

A Improved Generation Adversarial Network Image Inpainting Method based on Attention Mechanism and Texture Structure Stratification

作　　者：张剑飞 张洒 ZHANG Jianfei;ZHANG Sa(School of Computer&Information Engineering,Heilongjiang University of Science&Technology,Harbin 150022,China)

机构地区：[1]黑龙江科技大学计算机与信息工程学院,哈尔滨150022

出　　处：《牡丹江师范学院学报（自然科学版）》2022年第4期22-28,共7页Journal of Mudanjiang Normal University：Natural Sciences Edition

基　　金：国家自然科学基金项目(61803148)。

摘　　要：构建一个以U-Net为基础的模型,通过引入注意力机制与纹理结构分层相融合的图像修复方法,在生成对抗网络模型基础上,引入通道注意力并结合多尺度卷积模块,将图像下采样提取特征分为纹理特征与结构特征,采用改进的Res2Net残差块并重构损失函数.实验结果表明,修复后图像的纹理和结构更统一,修复图像与原始图像在高级语义上更加接近.A model based on U-net is constructed,by introducing an image inpainting method that integrates attention mechanism and texture structure layering,on the basis of generating adversarial network model,channel attention is introduced and combined with a multi-scale convolution module to divide the image down sampling extraction features into texture features and structural features.The improved Res2Net residual block is adopted and the loss function is reconstructed.The experimental results show that the texture and structure of the inpainted image are more uniform,and the inpainted image is closer to the original image in high-level semantics.

关 键 词：图像修复 注意力机制 纹理结构 生成对抗网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]